Workflow automation
In short
Workflow automation is the computerised facilitation or automation of a business process, in whole or in part, by passing documents, information or tasks between participants according to set rules.
- Also called
- Automated workflow, Process automation, Workflow management
- Topic
- AI agents
- Level
- Beginner
The Workflow Management Coalition defined workflow as the computerised facilitation or automation of a business process, in whole or in part1. It concerns procedures in which documents, information or tasks pass between participants according to a defined set of rules to achieve or contribute to an overall business goal1. A workflow may be organised by people alone, but the Coalition's reference model is aimed at computerised support1. A process definition lists activity steps, with computer operations, human operations, or both, and rules for how the process moves forward1.
At run time, software creates and controls instances of that process, schedules the steps and calls the people and IT applications each step needs1. The core piece is the workflow engine, which creates and deletes processes, schedules activities and interacts with application tools or human resources1. The 1994 reference model names five interface areas: interchange of process-definition data, interoperability between workflow systems, interaction with IT applications, interaction with desktop user interfaces, and monitoring and metrics1. A single process can last from minutes to days or months1.
Google Cloud Workflows is one current product2. A workflow is a series of steps written in YAML or JSON2. By default, steps run one at a time, in order, until all the steps have run2. The definition can use a switch, a for loop, parallel steps, subworkflows and callbacks that wait for another service to call back2. It can retry HTTP calls with exponential back-offs, and a run can hold state, retry, poll or wait for up to a year2. A run starts from the console or command line, from client libraries or an HTTP POST, or on a schedule with Cloud Scheduler2. It can encode conditions, actions and human-in-the-loop events2. Examples include tracking an order from request to fulfilment, and resource requests with approvals2.
Anthropic, in a note published on 19 December 2024, uses workflow for a narrower case: language models and tools orchestrated through predefined code paths, as against agents that direct their own tool use3. In that note, workflows are the more predictable option for well-defined tasks3. The same page says much of the tooling it describes has changed since December 20243.
Example
Google Cloud Workflows gives tracking an order from request to fulfilment, and resource requests with approvals, as examples of encoding the steps in a business process2.
Why it matters
The reference model separates the rules of a business process from the applications that carry out each step, so a later change can be made in the process definition1.
Questions people ask
Is workflow automation the same as an AI agent?
No. The 1994 reference model is about passing work between people and IT applications under defined rules, and a step can be a human operation1. Anthropic's December 2024 note uses workflow for language models and tools on predefined code paths, and agent for models that direct their own tool use3.